| name | demo-seed |
| description | Generate realistic, screenshot-ready seed data for MemOS — a believable org with teams, projects, agents, OKRs, briefs, and a rich graph of facts/learnings/artifacts. Use before demos, screenshots, or recording the README Loom. |
MemOS Demo Seed
A portfolio project lives or dies on the demo. Empty tables screenshot badly. This skill populates a believable, dense, internally-consistent dataset so the dashboard looks alive and the provenance graph has depth.
Output: packages/api/src/db/seed.ts (idempotent, run via pnpm db:seed)
Build a coherent narrative (not random noise)
Seed ONE believable org so everything connects:
- 1 org, 2 teams (e.g.
team.ml-platform, team.growth).
- 3 projects across them (e.g.
project.rag-search, project.onboarding-funnel, project.cost-optimization). Mark one okrs_required=true.
- 5-6 agents with varied
trust_score (0.62 → 0.98) and last_checkin_at spread across the last week. Include one low-trust agent (for the leaderboard contrast) and one platform "critic" agent.
- Per project: 1-2 OKRs, each with 3-6 milestones, mixed
status (some achieved, some pending), realistic metric_current vs metric_target so rollup bars show partial progress. Use varied units (percent, cents per piece, seconds, USD) and both directions (up/down).
- 8-12 workflow runs across agents, mixed
status (mostly complete, a couple open, one failed).
- 15-25 facts and 15-25 learnings attached to those runs — with REAL-sounding claims (draw domain flavor from RAG, onboarding, infra). Most medium/high ones cite an artifact; include a couple low-confidence unbacked ones.
applies_to uses problem-domain tags only.
- 8-10 artifacts (kind: log/screenshot/benchmark) referenced by the facts/learnings.
- 4-6 briefs: a mix of org-scoped (compliance reminders), team-scoped, and one question-answer brief. Leave 1-2 unacked so the dashboard shows the "pending briefs" state.
- 2-3 open questions and a couple of
choices (one closed with outcome, one open/stale for the loop-close critic to flag).
Make it screenshot-perfect
- Timestamps spread realistically (not all
now()), so the activity feed has a believable timeline.
- At least one learning with a high
reuse_count and reuse_success_count (the "compounding capital" story).
- One rich provenance chain: a high-reuse learning → its artifact → its run → its OKR → its agent, so the React Flow graph has a deep, clickable path to show off.
- UTF-8 content with a
≤, a —, and an emoji somewhere to prove encoding is clean.
Idempotency
Truncate-then-insert, or upsert by stable IDs, so re-running pnpm db:seed is safe.
Done when
pnpm db:seed runs clean, the dashboard's OKR tree / activity feed / provenance graph / trust leaderboard all look full and believable, and there's at least one deep provenance chain worth screenshotting.
